Add files using upload-large-folder tool

EVALUATION_ARCHITECTURE.md

@@ -1,73 +1,94 @@
 # CounselBench evaluation architecture
 
 CounselBench uses the same functional loop as enterprise agent benchmarks:
-isolated task state, an agent process, captured trajectories, and an independent
-grading step. Harbor provides orchestration; a pinned filesystem MCP provides
-the sandbox; a deterministic legal verifier grades investigation, decision,
-state, and answer.
+isolated provider state, an agent process, captured trajectories, and an
+independent deterministic grader. Harbor provides orchestration; the task pack
+serves a closed multi-provider MCP; CounselScore grades investigation,
+reasoning, provider state, communication, readback, and containment.
 
 ## Runtime flow
 
 ```text
-high-level employee request + 96 immutable source records
-                         |
-                         v
-              filesystem MCP sandbox
-                         |
-                         v
-  identity -> authority -> operations -> approval/revision
-                         |
-              +----------+----------+
-              |                     |
-       supported action        evidence hold
-              +----------+----------+
-                         |
-        decision + matter-register state + advice
-                         |
-             exact post-write readback
-                         |
-                         v
-             hidden deterministic verifier
+high-level employee question + 97 provider-bound assets
+                           |
+                           v
+          Clio Manage | Gmail | Drive | Slack sandbox MCP
+                           |
+                           v
+     identity -> authority -> current fact -> approval/capacity
+                           |
+                +----------+----------+
+                |                     |
+         supported action        evidence hold
+                +----------+----------+
+                           |
+       Clio register + Clio note + task-native communication
+                           |
+              native provider readback after each write
+                           |
+                           v
+              hidden deterministic CounselScore verifier
 ```
 
-The agent receives neither the gold output nor the verifier token. It can reach
-only `/workspace/documents` and `/workspace/output` through six tool
-contracts pinned to the official MCP filesystem server. Hidden world
-specification files are outside the allowlisted roots.
+The agent receives neither gold state nor the verifier capability token.
+Provider tools expose only task-scoped synthetic resources. Hidden world specs,
+expected values, and reference trajectories stay outside the agent boundary.
 
-Each task has twelve portfolio items, 97 files in twelve folders and nine
-native formats, 56–67 required evidence reads, 4–12 custody checks,
-task-specific searches, three writes, and three readbacks. Reference
-trajectories range from 76 to 93 successful calls. The verifier checks causal
-prerequisites and outcomes, not one exact call order.
+## Provider fidelity
+
+Each public tool maps to one documented operation:
+
+- Clio Manage v4 matters and notes (`GET`, `POST`, and `PATCH`);
+- Gmail v1 messages (`list`, `get`, and `send`);
+- Google Drive v3 files and comments (`list`, `get`, and `create`); and
+- Slack Web API search, conversation, and post operations.
+
+The sandbox owns deterministic implementation and synthetic state. It does not
+rename a business outcome into a fake API. Request nesting, identifiers,
+resource paths, methods, read/write annotations, and provider response shapes
+remain provider-shaped. A release-time contract audit rejects pseudo-tools,
+missing upstream metadata, permissive schemas, and incorrect mutation flags.
+
+## Evidence and state model
+
+Every task has twelve portfolio items, 97 assets in twelve folders and nine
+native formats, and 58–86 material reads. Each asset has an immutable evidence
+ID, provider record ID, digest, revision timestamp, source role, and material
+flag. The remaining assets are inspectable supporting references rather than
+secretly mandatory padding.
+
+Reference trajectories contain 69–97 successful calls. All 100 raw tool-name
+sequences and semantic action graphs are distinct. The verifier checks causal
+prerequisites and resulting provider state; it does not require one arbitrary
+global order among independent evidence reads.
 
 ## Archipelago/APEX mapping
 
 | Layer | CounselBench / Blobfish | Archipelago / APEX analogue |
 |---|---|---|
-| Task state | Immutable source room, writable output state, per-trial trace | Environment image plus initial snapshot |
-| Tool gateway | Pinned official filesystem MCP contract with closed offline implementation | Environment-server APIs and task-specific tools |
+| Task state | Immutable provider records plus isolated mutable Clio/collaboration state | Environment image plus initial snapshot |
+| Tool gateway | Documented provider operations over a closed sandbox MCP | Environment-server APIs and task-specific tools |
 | Agent execution | Harbor agent container | Agent service and selectable implementation |
-| Trajectory | Complete MCP trace plus agent and verifier artifacts | Trajectory, final snapshot, prompts, and agent logs |
-| Grading | Deterministic causal, branch, state-diff, containment, and readback criteria | Independent grading service with composable graders |
-| Reproducibility | 100 oracle trials, 1,000 negative controls, 100 exact replays, contract conformance | Containerized reruns against frozen environment and grader |
+| Trajectory | Full MCP call/result digest trace plus agent and verifier artifacts | Trajectory, final snapshot, prompts, and agent logs |
+| Grading | CounselScore semantic milestones backed by atomic evidence, state, containment, and readback checks | Independent grading service with composable graders |
+| Reproducibility | 100 oracle trials, 1,300 negative controls, 100 exact replays, provider contract audit | Containerized reruns against frozen environment and grader |
 
-The similarity is functional, not a claim that implementations are
+The similarity is functional, not a claim that the implementations are
 interchangeable. Archipelago separates environment, agents, and grading into
-services. CounselBench delegates orchestration to Harbor, embeds the MCP world
-inside every task pack, and keeps verification behind a capability token.
+services. CounselBench delegates orchestration to Harbor, embeds its MCP world
+inside each task pack, and protects verification with a capability token.
 
-## Why deterministic grading works here
+## Why deterministic grading works
 
 The generator creates raw facts and independently recomputes every branch from
-four controls. The evidence does not carry the computed disposition. The
-verifier can therefore check exact IDs, revisions, values, source paths,
-owners, deadlines, state rows, and trace ordering without a model judge.
+four controls. Evidence does not carry the computed disposition. The verifier
+can therefore check exact IDs, revisions, values, source roles, owners,
+deadlines, task-scoped state, and trace ordering without a model judge.
 
-This design intentionally limits the scored claim. Open-ended legal
-persuasiveness is not graded. Whether the agent found the right records,
-applied the released decision rule, changed only the task-scoped state, and
-reported the exact supported result is graded.
+The scored claim is intentionally bounded. Open-ended legal persuasiveness is
+not graded. Whether the agent found the right records, applied the authored
+decision rule, changed only authorized provider state, communicated the exact
+supported result, and verified committed state is graded.
 
 ## Public references
 
@@ -75,4 +96,7 @@ reported the exact supported result is graded.
 - [APEX Accounting benchmark](https://www.mercor.com/apex/apex-accounting-leaderboard/)
 - [APEX Agents leaderboard](https://www.mercor.com/apex/apex-agents-leaderboard/)
 - [Harbor Framework](https://harborframework.com/)
-- [Model Context Protocol servers](https://github.com/modelcontextprotocol/servers)
+- [Clio Manage API](https://docs.developers.clio.com/api-docs/clio-manage/)
+- [Gmail API](https://developers.google.com/workspace/gmail/api/reference/rest/v1/users)
+- [Google Drive API](https://developers.google.com/drive/api/reference/rest/v3)
+- [Slack Web API](https://api.slack.com/methods)

README.md

@@ -23,15 +23,16 @@ configs:
 
 # CounselBench-100
 
-CounselBench-100 v3.1.0 is a synthetic legal-work benchmark with 100
+CounselBench-100 v3.2.0 is a synthetic legal-work benchmark with 100
 authored matters across ten practice workflows. Every task has a natural employee
-request, a 97-file evidence room, twelve portfolio decisions, 5–9 supported
-actions, 3–7 evidence holds, and a distinct 76–93-call filesystem MCP trajectory.
+request, a 97-asset evidence room, twelve portfolio decisions, 5–9 supported
+actions, 3–7 evidence holds, and a distinct deep multi-provider MCP trajectory.
 
 The answer is not preclassified in the evidence. Each portfolio item requires an
 immutable identity join, an operative-authority and revision lookup, a current-state
-comparison, and an effective approval/owner-capacity check. The agent then creates
-an exact matter-register row and reads all three final work products back.
+comparison, and an effective approval/owner-capacity check across Clio, Gmail,
+Drive, and Slack. The agent then commits native provider state and reads each
+changed record back.
 
 ## Public release
 
@@ -42,33 +43,33 @@ an exact matter-register row and reads all three final work products back.
 
 ## Included files
 
-- `data/tasks.jsonl`: prompt, context paths, public causal rubric, gold output, and metadata.
+- `data/tasks.jsonl`: prompt, provider-bound assets, 14 semantic milestones, gold output, and metadata.
 - `task_files/`: 9,700 unique files across Markdown, TXT, EML, CSV, JSON, XML, HTML, PDF, and XLSX.
 - `world/`: pinned offline MCP implementation and hidden deterministic verifier.
 - `trajectories/`: 100 solvability traces; these are excluded from model ranking.
 - `reports/`: exact-version build, qualification, and conformance evidence.
 - `SCORING.md`: causal, branch, state, containment, and readback contract.
 
-## Measured v3.1.0 release gates
+## Measured v3.2.0 release gates
 
 | Gate | Measured |
 |---|---:|
 | Tasks | 100 |
 | Agent-visible files | 9,700 unique; all nine native formats parse |
-| Required evidence reads | 56–67 per task |
-| Reference MCP calls | 76–93 per task |
-| Raw tool sequences | 100 distinct |
-| Semantic action graphs | 100 distinct; maximum pair match 0.169697 |
-| Prompt maximum 5-shingle Jaccard | 0.386139 |
+| Required evidence reads | 58–86 per task |
+| Reference MCP calls | 69–97 per task |
+| Raw tool sequences | 100 distinct; maximum pair match 0.807453 |
+| Semantic action graphs | 100 distinct; maximum pair match 0.165517 |
+| Prompt maximum 5-shingle Jaccard | 0.267974 |
 | Oracle and deterministic replay | 100/100 each |
-| Ten negative controls | 1,000/1,000 rejected |
+| Thirteen negative controls | 1,300/1,300 rejected |
 
-No v2 or partial model score is carried onto v3. A leaderboard row is published
+No older or partial model score is carried onto v3.2. A leaderboard row is published
 only after one model runs all 100 tasks on this exact release.
 
 ## MCP fidelity and licenses
 
-The world mirrors an allowlisted subset of `@modelcontextprotocol/server-filesystem@2026.7.10`
-at upstream commit `9a96ea6e5913736f92b88345bf51caeaaa8e719f`. Live contract and behavior conformance is
-versioned separately. Synthetic data is CC-BY-4.0; benchmark code is
+The world maps each allowlisted tool to a documented Clio Manage v4, Gmail v1,
+Google Drive v3, or Slack Web API operation. Contract metadata and source links
+ship with the release. Synthetic data is CC-BY-4.0; benchmark code is
 Apache-2.0. Every entity, person, event, amount, and address is fictitious.

SCORING.md

@@ -1,62 +1,75 @@
 # CounselBench-100 scoring contract
 
-CounselBench-100 v3.1 reports deterministic criterion credit and a separate
-strict pass. The grader makes no model, network, clock, locale, or random call.
+CounselBench-100 v3.2 reports CounselScore from fourteen semantic employee
+outcomes totaling 100 points. The public rubric describes what a competent
+employee must accomplish. The hidden deterministic verifier retains the exact
+field-, evidence-, state-, and trace-level checks needed to reproduce that
+score. It makes no model, network, clock, locale, or random call.
 
-## What v3 measures
+## What the task requires
 
-The release no longer rewards copying answer-shaped metadata. Source records do
-not label a finding, identify a primary or corroborating record, assign
-severity, name a remediation owner, or disclose the selected option. For each
-of twelve portfolio keys, the agent must establish four independent controls:
+The source records do not label findings, identify a primary record, assign
+severity or ownership, or reveal the selected option. For each of twelve
+portfolio keys, the agent must establish four independent controls:
 
 1. resolve the immutable entity rather than match a display name;
 2. locate the operative authority and current effective revision;
-3. compare the current operational fact to the governed trigger;
+3. compare the current operational fact to the governed trigger; and
 4. verify effective approval and an active owner with capacity.
 
 All four controls produce a supported action. A failed control produces a
-specific evidence hold. Every task contains both branches.
+specific evidence hold. Every task contains both branches. Financial,
+population, chronology, correspondence, and independent counterrecords become
+material only where the task-specific branch needs them.
 
-The employee-facing request stays high level. A neutral work-product protocol
-in the evidence room defines the output shapes and presents all three candidate
-approaches without marking an answer.
+The employee request remains high level. A neutral protocol in the matter
+record describes the firm’s work-product controls and all candidate approaches
+without selecting an answer or prescribing a call sequence.
 
-## Scored categories
+## CounselScore milestones
 
-| Category | Weight | What is checked |
+| Milestone | Points | Deterministic outcome |
 |---|---:|---|
-| Investigation | 25% | Exact required reads, custody checks, discovery/search, and all evidence preceding the first write |
-| Decision | 35% | Selected option, alternatives, exact action and hold populations, branch facts, citations, owners, and deadlines |
-| State | 25% | One exact matter-register row per portfolio key, no collateral rows or fields, and write-scope containment |
-| Advice | 15% | Recommendation, causal rationale, every supported action, every hold, alternatives, and limitations |
+| Cross-system discovery | 6 | The exact matter and its source population are found in Clio, Gmail, Drive, and Slack |
+| Immutable identity | 9 | All portfolio records are joined by immutable identifiers; collisions remain holds |
+| Operative authority | 9 | The controlling authority and effective revision are used |
+| Current operations | 9 | Current events, values, and populations are compared with the applicable trigger |
+| Approval and capacity | 7 | Effective authority, active ownership, and capacity are established |
+| Impact and counterrecords | 6 | Task-specific workbook, correspondence, chronology, and counterrecords are correlated |
+| Option selection | 10 | The supported approach and both alternatives are correctly distinguished |
+| Branch classification | 10 | The exact supported-action and evidence-hold populations are derived |
+| Action details | 8 | Exact issues, identities, severity, owners, dates, determinations, actions, gaps, and citations are present |
+| Clio matter register | 8 | The exact twelve-row register is patched on the isolated matter |
+| Clio decision note | 6 | The decision-ready note contains the complete supported result and limits |
+| Collaboration state | 4 | After core Clio state is committed, the scoped completion message is committed to the established Gmail, Drive, or Slack channel |
+| Provider readback | 4 | Every changed provider record is re-opened after its final write |
+| Containment | 4 | All material evidence precedes mutation; the exact mutation set succeeds with no rejected write |
 
-The public rubric has 82 task-specific descriptions: global causal gates plus
-identity, authority, operations, approval, branch, and register checks for each
-portfolio key. The executable verifier expands those descriptions into
-field-level decision and state criteria.
+Each milestone is task-specific: its public description names the matter
+question, selected option, exact supported and held populations, and the native
+completion channel. Internally, every atomic check is assigned to exactly one
+milestone. Partial score is the sum of within-milestone atomic completion times
+the milestone weight. Strict pass requires all 100 points and every atomic
+check.
 
-## Ordering, containment, and readback
+## Native state and trace proof
 
 The trace—not a claimed narrative—proves the workflow:
 
-- all 56–67 required records must be read in full before the first write;
-- designated metadata checks and task-specific format searches must precede
-  mutation;
-- only `decision.json`, `matter-register.json`, and `advice.md` may be
-  written;
-- each final file digest must match the last MCP write;
-- each final file must be read back after its last write.
-
-## Reward caps and strict pass
-
-- Missing or out-of-scope state writes cap reward at 0.20.
-- Missing post-write verification caps reward at 0.35.
-- Incomplete or late required investigation caps reward at 0.49.
-- Another incomplete procedure caps reward at 0.59.
-- Strict pass requires every procedure, decision, state, and advice criterion.
-
-Exact decision and register equality are graded state properties, not hidden
-string checks in the prompt. All outcomes and category scores are included in
-the verifier report, so a published value can be reproduced without an
-evaluator model.
+- all 58–86 material provider records must be opened before the first mutation;
+- Clio, Gmail, Drive, and Slack discovery must cover the isolated matter;
+- the only successful writes are the exact Clio matter update, exact Clio note
+  creation, and the task-native completion communication;
+- failed exploratory reads are permitted and do not erase a correct outcome;
+- any rejected state-changing call fails containment;
+- duplicate or out-of-scope writes fail the exact mutation-set contract; and
+- the matter register and decision note must be committed before the completion
+  communication; and
+- Clio state and the completion communication must be read back after their
+  corresponding final writes, with response hashes matching current state.
+
+The register, decision note, and communication are graded as provider state,
+not as local files or hidden strings in the prompt. Verifier reports include
+the fourteen milestones, every atomic result, partial credit, state diagnostics,
+and a report digest so a published value can be audited without an evaluator
+model.